[Pacemaker] [PATCH]Bug 2567 - crm resource migrate should support an optional "role" parameter

[Holger Teutsch]

On Mon, 2011-04-04 at 15:24 +0200, Andrew Beekhof wrote:
> On Mon, Apr 4, 2011 at 2:43 PM, Holger Teutsch <holger.teutsch at web.de> wrote:
> > On Mon, 2011-04-04 at 11:05 +0200, Andrew Beekhof wrote:
> >> On Sat, Mar 19, 2011 at 11:55 AM, Holger Teutsch <holger.teutsch at web.de> wrote:
> >> > Hi Dejan,
> >> >
> >> > On Fri, 2011-03-18 at 14:24 +0100, Dejan Muhamedagic wrote:
> >> >> Hi,
> >> >>
> >> >> On Fri, Mar 18, 2011 at 12:21:40PM +0100, Holger Teutsch wrote:
> >> >> > Hi,
> >> >> > I would like to submit 2 patches of an initial implementation for
> >> >> > discussion.
> >> > ..
> >> >> > To recall:
> >> >> >
> >> >> > crm_resource --move resource
> >> >> > creates a "standby" rule that moves the resource off the currently
> >> >> > active node
> >> >> >
> >> >> > while
> >> >> >
> >> >> > crm_resource --move resource --node newnode
> >> >> > creates a "prefer" rule that moves the resource to the new node.
> >> >> >
> >> >> > When dealing with clones and masters the behavior was random as the code
> >> >> > only considers the node where the first instance of the clone was
> >> >> > started.
> >> >> >
> >> >> > The new code behaves consistently for the master role of an m/s
> >> >> > resource. The options "--master" and "rsc:master" are somewhat redundant
> >> >> > as a "slave" move is not supported. Currently it's more an
> >> >> > acknowledgement of the user.
> >> >> >
> >> >> > On the other hand it is desirable (and was requested several times on
> >> >> > the ML) to stop a single resource instance of a clone or master on a
> >> >> > specific node.
> >> >> >
> >> >> > Should that be implemented by something like
> >> >> >
> >> >> > "crm_resource --move-off --resource myresource --node devel2" ?
> >> >> >
> >> >> > or should
> >> >> >
> >> >> > crm_resource refuse to work on clones
> >> >> >
> >> >> > and/or should moving the master role be the default for m/s resources
> >> >> > and the "--master" option discarded ?
> >> >>
> >> >> I think that we also need to consider the case when clone-max is
> >> >> less than the number of nodes. If I understood correctly what you
> >> >> were saying. So, all of move slave and move master and move clone
> >> >> should be possible.
> >> >>
> >> >
> >> > I think the following use cases cover what can be done with such kind of
> >> > interface:
> >> >
> >> > crm_resource --moveoff --resource myresource --node mynode
> >> >   -> all resource variants: check whether active on mynode, then create standby constraint
> >> >
> >> > crm_resource --move --resource myresource
> >> >   -> primitive/group: convert to --moveoff --node `current_node`
> >> >   -> clone/master: refused
> >> >
> >> > crm_resource --move --resource myresource --node mynode
> >> >  -> primitive/group: create prefer constraint
> >> >  -> clone/master: refused
> >>
> >> Not sure this needs to be refused.
> >
> > I see the problem that the node where the resource instance should be
> > moved off had to be specified as well to get predictable behavior.
> >
> > Consider a a 2 way clone on a 3 node cluster.
> > If the clone is active on A and B what should
> >
> > crm_resource --move --resource myClone --node C
> >
> > do ?
> 
> I would expect it to create the +inf constraint for C but no
> contraint(s) for the current location(s)

You are right. These are different and valid use cases.

crm_resource --move --resource myClone --node C
   -> I want an instance on C, regardless where it is moved off

crm_resource --move-off --resource myClone --node C
   -> I want the instance moved off C, regardless where it is moved on

I tried them out with a reimplementation of the patch on a 3 node
cluster with a resource with clone-max=2. The behavior appears logical
(at least to me 8-) ).

> 
> > This would require an additional --from-node or similar.
> >
> >> Other than that the proposal looks sane.
> >>
> >> My first thought was to make --move behave like --move-off if the
> >> resource is a clone or /ms, but since the semantics are the exact
> >> opposite, that might introduce introduce more problems than it solves.
> >
> > That was my perception as well.
> >
> >>
> >> Does the original crm_resource patch implement this?
> >
> > No, I will submit an updated version later this week.
> >
> > - holger
> >
> >
> > _______________________________________________
> > Pacemaker mailing list: Pacemaker at oss.clusterlabs.org
> > http://oss.clusterlabs.org/mailman/listinfo/pacemaker
> >
> > Project Home: http://www.clusterlabs.org
> > Getting started: http://www.clusterlabs.org/doc/Cluster_from_Scratch.pdf
> > Bugs: http://developerbugs.linux-foundation.org/enter_bug.cgi?product=Pacemaker
> >